光学学报, 2018, 38 (4): 0411009, 网络出版: 2018-07-10
基于改进区域项CV模型的金相图像分割 下载: 686次
Segmentation of Metallographic Image Based on Improved CV Model Integrated with Local Fitting Term
成像系统 金相图像分割 Chan-Vese模型 区域项 倒数交叉熵 最大绝对中位差 imaging systems segmentation of metallographic image Chan-Vese model local fitting term reciprocal cross entropy maximum absolute median difference
摘要
为了解决传统Chan-Vese(CV)模型难以快速、精确提取金相晶粒的问题,提出一种基于改进区域项CV模型的金相图像分割方法。该方法利用倒数交叉熵阈值选取准则函数替代传统CV模型中能量函数的区域项,构造新的水平集模型。改进模型能够使分割前后图像的倒数交叉熵达到最小,更精确地分割噪声影响严重且局部灰度变化较大的金相图像;考虑到倒数交叉熵计算会增加算法复杂度,通过引入最大绝对中位差,自适应调整曲线内外的能量权重加速曲线的演化,添加距离规范项以避免水平集函数的重新初始化,加速模型的收敛。实验结果表明,与多种模型相比,改进模型在分割结果和分割效率方面均具有明显优势。
Abstract
In order to solve the problem that traditional Chan-Vese (CV) model is difficult to extract metallographic grains quickly and accurately, the metallographic image segmentation method based on improved CV model integrated with local fitting term is proposed. We use the reciprocal cross entropy threshold segmentation rule to replace the regional term of the energy function in the traditional CV model and construct a new level set model. The proposed model can minimize the reciprocal cross entropy between original and segmented image, and accurately segment the metallographic images with more noises and larger local gray scale. In addition, Taking that the reciprocal cross entropy will increase algorithm’s computational complexity into account, the maximum absolute median difference is adopted to adjust energy weight inside and outside the curve to accelerate curve evolution. The distance regularized term is introduced to avoid initialing level set function, and accelerate the model convergence. Experimental results show that comparing with other traditional CV models, the proposed model has obvious advantages both in segmentation result and efficiency.
倪康, 吴一全, 庚嵩. 基于改进区域项CV模型的金相图像分割[J]. 光学学报, 2018, 38(4): 0411009. Kang Ni, Yiquan Wu, Song Geng. Segmentation of Metallographic Image Based on Improved CV Model Integrated with Local Fitting Term[J]. Acta Optica Sinica, 2018, 38(4): 0411009.